This page may contain stale information. Last updated: 2026-08-15
Definition
Homomorphic encryption (HE/FHE) allows computation on encrypted data without decrypting it on the server — enabling private AI inference where the provider never sees plaintext inputs. Distinct from end-to-end encryption (server cannot compute at all).
Key Points
- 2026-08-14: google open-sourced heir-compiler — MLIR toolchain converting pretrained models to encrypted-input inference
- Four demo apps: recommendations, fraud detection, Kitsune anomaly detection, hotword detection
- Hardware partners: Belfort, Niobium, Cornami, Optalysys
-
Production latency overhead typically 100x–1000x+ vs plaintext on CPU; suited for regulated niches (healthcare, finance), not general LLM chat